temporaldynamicsofmemory-guidedcognitivecontrol ... · these findings were cross-validated using...

Behavioral/Cognitive

Temporal Dynamics of Memory-guided Cognitive Controland Generalization of Control via Overlapping AssociativeMemories

X Jiefeng Jiang,1 X Inês Bramão,2 Anna Khazenzon,1 Shao-Fang Wang,1 X Mikael Johansson,2and X Anthony D. Wagner1,31Department of Psychology, Stanford University, Stanford, California 94305, 2Department of Psychology, Lund University, Lund, Sweden, SE-221 00,and 3Wu Tsai Neurosciences Institute, Stanford University, Stanford, California 94305

Goal-directed behavior can benefit from proactive adjustments of cognitive control that occur in anticipation of forthcoming cognitivecontrol demands (CCD). Predictions of forthcoming CCD are thought to depend on learning and memory in two ways: First, throughdirect experience, associative encoding may link previously experienced CCD to its triggering item, such that subsequent encounters withthe item serve to cue retrieval of (i.e., predict) the associated CCD. Second, in the absence of direct experience, pattern completion andmnemonic integration mechanisms may allow CCD to be generalized from its associated item to other items related in memory. Whileextant behavioral evidence documents both types of CCD prediction, the neurocognitive mechanisms giving rise to these predictionsremain largely unexplored. Here, we tested two hypotheses: (1) memory-guided predictions about CCD precede control adjustments dueto the actual CCD required; and (2) generalization of CCD can be accomplished through integration mechanisms that link partiallyoverlapping CCD-item and item-item associations in memory. Supporting these hypotheses, the temporal dynamics of theta and alphapower in human electroencephalography data (n � 43, 26 females) revealed that an associative CCD effect emerges earlier than interac-tion effects involving actual CCD. Furthermore, generalization of CCD from one item (X) to another item (Y) was predicted by a decreasein alpha power following the presentation of the X-Y pair. These findings advance understanding of the mechanisms underlying memory-guided adjustments of cognitive control.

Key words: associative memory; cognitive control; EEG; generalization

IntroductionCognitive control refers to a collection of neurocognitive func-tions that align behavior with internal goals through top-down

modulations on neural information processing, and hence plays akey role in adaptive behavior (Miller and Cohen, 2001; Waskomet al., 2014; Egner, 2017). One key feature of cognitive control isthat it adjusts to meet the cognitive control demand (CCD) of thepresent environment (Botvinick et al., 1999; Kerns et al., 2004).

Received Aug. 1, 2019; revised Jan. 29, 2020; accepted Jan. 29, 2020.Author contributions: J.J., A.K., and A.D.W. designed research; J.J., A.K., and S.-F.W. performed research; J.J., I.B.,

A.K., S.-F.W., M.J., and A.D.W. analyzed data; J.J. wrote the first draft of the paper; J.J., I.B., A.K., S.-F.W., M.J., andA.D.W. edited the paper; J.J., I.B., M.J., and A.D.W. wrote the paper.

This work was supported by grants from the National Institute on Aging F32AG056080 to J.J., and R21AG058111to A.D.W., and a Marcus and Amelia Wallenberg Foundation Award MAW2015.0043 to M.J. and A.D.W. We thank Dr.Russell Poldrack for helpful comments on an earlier version of this manuscript.

The authors declare no competing financial interests.Correspondence should be addressed to Jiefeng Jiang at [email protected]://doi.org/10.1523/JNEUROSCI.1869-19.2020

Copyright © 2020 the authors

Significance Statement

Cognitive control adaptively regulates information processing to align with task goals. Experience-based expectations enableadjustments of control, leading to improved performance when expectations match the actual control demand required. UsingEEG, we demonstrate that memory for past cognitive control demand proactively guides the allocation of cognitive control,preceding adjustments of control triggered by the demands of the present environment. Furthermore, we demonstrate thatlearned cognitive control demands can be generalized through mnemonic integration processes, enabling the spread of expecta-tions about cognitive control demands to items associated in memory. We reveal that this generalization is linked to decreasedalpha oscillation in medial frontal channels. Collectively, these findings provide new insights into how memory-control interac-tions facilitate goal-directed behavior.

The Journal of Neuroscience, Month XX, 2020 • 40(XX):XXXX–XXXX • 1

mailto:[email protected]

For example, a driver reacts to worsened driving conditions byflexibly increasing attention to the road and suppressing irrele-vant information and behavior. A second key feature of control isthat adjustments can be proactive, preparing an individual tomeet anticipated imminent demands (Braver, 2012). Emergingdata indicate that proactive adaptations are often guided by asso-ciative learning and memory (Egner, 2014; Abrahamse et al.,2016; Braem et al., 2019; Chiu and Egner, 2019). For instance, anexperience-based association between a busy overpass and highCCD could lead to memory-guided proactive engagement ofcontrol the next time one approaches the overpass. The modula-tion of item-CCD association on cognitive control has been dem-onstrated in human behavior (Jacoby et al., 2003; Bugg et al.,2011), with the encoding of item-CCD associations modeled bytemporal difference learning across trials (Chiu et al., 2017).

A central open question is as follows: how does memory guideadjustments of cognitive control to align control with imminentCCDs? Intuitively, learning the CCD associated with an itemshould allow an organism to proactively adapt cognitive controlto the predicted CCD before the actual demands are detected. Todate, this idea has been modeled in computational simulations(Blais et al., 2007; Verguts and Notebaert, 2008), yet empiricaltests are scarce. Thus, the first goal of this study is to test thishypothesis by examining the temporal dynamics of association-guided cognitive control.

While the acquisition of item-CCD associations depends, inpart, on striatal mechanisms (Chiu et al., 2017), learning oftenoccurs in multiple neural systems that support distinct memorytypes and processes (Poldrack and Packard, 2003; Kumaran et al.,2016). Hippocampal-dependent mechanisms may support thegeneralization of expectations about control to related items inmemory. For example, the high CCD associated with the busyoverpass may be generalized to the roads near the overpass with-out directly experiencing high CCD on those roads. Indeed, thegeneralization of CCD has been documented in human behavior(Crump and Milliken, 2009; King et al., 2012; Weidler and Bugg,2016; Surrey et al., 2017; Bejjani et al., 2018), although the neu-rocognitive mechanisms remain poorly understood. As a secondgoal, we tested the hypothesis that the generalization of CCD canbe achieved through integrative encoding (Shohamy and Wag-ner, 2008; Kuhl et al., 2010), wherein partially overlapping asso-ciations (e.g., overpass-road and overpass-CCD) result in theformation of an integrated representation (e.g., overpass-road-CCD) that supports direct retrieval of CCD expectations for anitem (e.g., road as cue) that have been inherited from anotherassociated item (e.g., overpass).

To test these hypotheses, we leveraged the high temporal res-olution of EEG along with a learning and generalization para-digm. Similar to previous studies of generalization (Zeithamovaand Preston, 2010; Kuhl et al., 2011; Wimmer and Shohamy,2012; Zeithamova et al., 2012; Bejjani et al., 2018), the task con-sisted of three phases (Fig. 1): an association phase establishingtool–landmark associations, a training phase introducing tool–CCD associations, and a test phase assessing the generalization ofCCD from tools to landmarks. To preview the results, in thetraining phase EEG data, the observed temporal dynamics of neu-ral responses are consistent with associative-memory driven pro-active engagement of control that precedes further adjustmentsof control in response to the actual CCD required by the trial.These findings were cross-validated using the independent testphase EEG data. Moreover, the behavioral data at test and EEGdata during the association and test phases provide strong evi-dence of generalization of CCD via associative memory.

Materials and MethodsSubjects. Fifty-three subjects gave informed written consent, in accor-dance with procedures approved by the Stanford University InstitutionalReview board. Data from 4 subjects were excluded due to low behavioralperformance (accuracy was �3 SDs lower than the group median) in atleast one experimental condition of at least one of the three phases (seebelow) (Leys et al., 2013). Data from 6 additional subjects were excludeddue to excessive EEG artifacts. The final sample consisted of 43 partici-pants (18 –29 years old, mean � 22.1 years; 26 females) with normal orcorrected-to-normal vision and no history of psychiatric or neurologicaldisorders.

Stimuli and experimental design. The stimuli consisted of eight colorimages: four tools and four landmarks (Fig. 1A). The images were pre-sented on a 23-inch LCD display at 60 Hz using Psychtoolbox 3 andcovered �7.7° of visual angle. The task consisted of three phases: anassociation phase, a training phase, and a test phase.

The association phase (Fig. 1A) aimed to elicit the incidental encodingof tool–landmark associations. To do so, the association phase com-prised 6 runs of 60 trials each. Each trial consisted of the pairing of aspecific tool followed by a specific landmark; the pairings were repeatedthroughout the association phase, creating four unique tool–landmarkassociations. The specific pairings were randomized across participants.Throughout, each image was displayed for 800 ms and the tool–land-mark images were separated by a uniformly jittered interstimulus inter-val (900 –1100 ms). To temporally separate the trials and promote theencoding of the tool–landmark associations, the intertrial intervals wereuniformly jittered between 2250 and 2750 ms (i.e., the intertrial intervalswere substantially longer than the interstimulus intervals). Participantswere not instructed to intentionally encode the tool–landmark associa-tions. Instead, to ensure that participants attended to the images, theirtask was to press a response button using their right index finger when-ever the encountered image was inverted. A tool image was defined asinverted when its handle was shown in the bottom half of the image;landmarks were inverted when their base was above their roof. Therewere four presentations of inverted tools and four inverted landmarks ineach run.

The goal of the training phase was to associate each of the four toolswith either a high or low CCD. To this end, participants performed avariant of the Stroop task (Fig. 1B). On each trial, a compound stimulus,consisting of a tool image (target) and a superimposed tool name (dis-tractor), was presented for 800 ms. The participants were required toidentify the tool in the image by pressing a response button while tryingto ignore the tool name. Participants used four fingers of the same handto separately respond to the four tools. The response mapping wascounterbalanced across participants. Trials were separated by uniformlyjittered intertrial intervals (2700 –3100 ms). It is well established that,compared with congruent trials in which the target and distractor lead tothe same response, incongruent trials require higher CCD to resolve theresponse conflict between the target and distractor (Cohen et al., 1990;Botvinick et al., 2001).

The CCD associated with each of the four tools was varied using anitem-specific proportion of conflict (ISPC) manipulation during train-ing. Specifically, two tools were associated with high CCD (denoted asTH1 and TH2) by being presented in incongruent trials 75% of the time(i.e., ISPC � 75%), whereas the other two tools were associated with lowCCD (denoted as TL1 and TL2) by being presented in incongruent trials25% of the time (i.e., ISPC � 25%). The training phase comprised 6 runsof 48 trials each, with 12 trials per tool image per run. As such, themanipulations resulted in a 2 (associative CCD, manipulated by ISPC) �2 (actual CCD, manipulated by congruency) factorial design. To fosterintegration (Zeithamova and Preston, 2010), the association and trainingphase runs were interleaved in sets of 2 (Fig. 1D); as detailed below, weinvestigated how neural activity in the association phase changed afterexposure to the tool–CCD associations in the training phase.

A final test phase was used to assess the generalization of CCD fromtool–CCD associations (established in the training runs) to landmark–CCD associations mediated through the tool–landmark associations (in-duced in the association phase). The task in the test phase was similar to

2 • J. Neurosci., Month XX, 2020 • 40(XX):XXXX–XXXX Jiang et al. • Dynamics of Memory-guided Control and Transfer

the task in the training phase. Participants were required to identify theimage of the landmark while trying to ignore the word label superim-posed on the image (Fig. 1C). The trial structure and image presentationtimes were identical to the training phase. To avoid any potential con-found due to the overlap in stimulus–response mappings, participantsresponded using the other hand than the one used in the training phase.Across the 4 test runs, two landmarks (LL2 and LH2) were presented inthe same ISPC as their associated tool. Crucially, the other two landmarks(LL1 and LH1) were presented in a neutral (50%) ISPC and were used totest the generalization of CCD without the potential confound of expe-riencing a biased (low or high) ISPC across the test phase. As such, anyISPC effects for these landmarks must be inherited from their associatedtools. Having biased landmarks (e.g., LL2 and LH2) is not necessary forgeneralization to occur (Bejjani et al., 2018).

Behavioral analysis. To test whether the participants were engagedduring the association phase, we calculated the hit rate (responding whenthe image was inverted) and overall accuracy (correctly making/with-holding a response based on task instructions).

In the training phase, we analyzed accuracy and response time (RT).Accuracy was analyzed using a repeated-measures ANOVA, includingthe factors, associated CCD (high/low) and congruency (congruent/in-congruent). RTs were analyzed using a model-based approach (Chiu etal., 2017) to assess learning of the tool–CCD associations. Specifically, thelearning of the CCD associated with a tool was modeled using a temporaldifference learning algorithm (Sutton and Barto, 2018) as follows:

Pi(t � 1) � (1 � �)Pi(t) � �C(t)

where C(t) represents the congruency (1 � incongruent; 0 � congruent)at Trial t; Pi quantifies the model-belief of the CCD associated with tool i;� is the learning rate that determines how strongly Pi is influenced byexperienced congruency. � was determined using a grid search (see be-low) and shared across all four tools. Given � and the trial sequenceexperienced by a participant, the model produces trial-by-trial estimatesof Pi (i.e., the probability that the forthcoming trial is incongruent) andPEi, which denotes the unsigned prediction error at Trial t (i.e., the

Figure 1. Experimental design of the three phases in the task. A, In the association phase, participants incidentally formed tool–landmark associations by viewing successively presented toolsfollowed by their respective associative landmarks. Participants responded to rare inverted images. B, In the training phase, participants performed a variant of the Stroop task, wherein theyidentified the tool and tried to ignore the superimposed word. This phase induced associations between tools and CCDs by manipulating how frequently each tool was used in incongruent trials. Eachstimulus is coded by its category (T, Tool; L, landmark), associated/transferred CCD (H, High; L, low), and a number to ensure uniqueness. For example, LL1 indicates a landmark whose associated toolwas paired with low CCD. Two tools (TL1 and TL2) were presented mostly in congruent trials (25% of ISPC), whereas the other two tools (TH1 and TH2) were used mostly in incongruent trials (75%of ISPC). C, In the test phase, participants performed the Stroop task but encountered the landmarks as stimuli. Two landmarks (LL1 and LH1) were presented using 50% ISPC. The other twolandmarks (LL2 and LH2) were presented using with the same ISPC as their associated tools in the training phase. D, There were 6 interleaved association and training runs, followed by 4 test runs.

Jiang et al. • Dynamics of Memory-guided Control and Transfer J. Neurosci., Month XX, 2020 • 40(XX):XXXX–XXXX • 3

absolute difference between Pi(t) and C(t)).These model estimates were used to explain thevariance in trialwise RTs as detailed below.

Trials accompanied by nuisance cognitiveprocesses (e.g., unsuccessful conflict resolutionand posterror slowing), such as error trials andposterror trials, were excluded from RT analy-ses. In addition, trials with RTs outside of thegrand median �2.5 SD range were excluded.For each participant, the remaining trialwiseRTs were regressed against a linear model with7 regressors (congruency, predicted controldemand, control prediction error, and 4 re-gressors representing each of the 4 tools). Theshared variance between the predicted CCDand the congruency regressors ranges from0.01 and 0.18 (0.14 � 0.01) across subjects.Low shared variance (e.g., �0.01) is possiblewith extreme learning rates (e.g., 1). Controlprediction error shared little variance withcongruency (�0.002 for all subjects) and pre-dicted CCD (� 0.001 for all subjects). Thelearning rate was determined by a grid search(range: 0 –1, step size � 0.01) that minimizedthe sum of squared errors of the model fit usingtrial-level RTs. The estimated coefficients inthe winning model were normalized using er-ror terms from model fitting and were thenpassed to group-level analyses, which usedone-sample t tests to examine whether themean of a coefficient was significantly differentfrom 0. The grid search does not constrain thesign of the estimated coefficients for the regres-sors and was thus orthogonal to group-levelanalysis.

For the test phase, the generalization of anassociated tool’s CCD to its landmark was an-alyzed using the items with neutral ISPCs (i.e., LL1 and LH1). Themodel-based analysis was not used because the focus of this analysis wasthe generalization of the CCD through tool–landmark associationsrather than the learning of a new CCD–landmark association from thetrial sequence in the test phase. Instead, we performed repeated-measures ANOVAs with the factors CCD of the associated tool (high/low) and congruency (congruent/incongruent) on accuracy and on themedian of RT in each condition.

EEG data acquisition and preprocessing. EEG data from 128-channelHydroCell Sensor Nets (Electrical Geodesics) were recorded at a sam-pling rate of 1000 Hz while participants performed the experiment. Animpedance threshold was set to 50k ohms and was checked approxi-mately every 12 min. EEG data were preprocessed using EEGlab (https://sccn.ucsd.edu/eeglab/index.php) and in-house MATLAB scripts. EEGrecordings were downsampled to 500 Hz and then went through anautomatic channel rejection procedure based on magnitude and varianceusing EEGlab. A high-pass filter of �0.1 Hz was applied to the remainingdata. For all three task phases, the continuous recorded data were dividedinto epochs of 1500 ms, ranging from �500 ms to 1000 ms poststimulusonset. Trial-level data went through the automatic epoch rejection ofEEGlab using the “all methods” option and default settings. Remainingtrials went through another manual epoch rejection process. Trials thatsurvived the rejection procedures were transformed using independentcomponent analysis for further manual rejection of components reflect-ing eye movements and noise. Independent component analysis-filtereddata were rereferenced to the average across all remaining channels.Missing channels were reconstructed using interpolation. Preprocesseddata were then used in both event-related potential (ERP) and time-frequency analyses. For ERP analysis, preprocessed EEG data were low-pass filtered (cutoff � 30 Hz), and the 200 ms before stimulus onset wasused for baseline correction. ERP data ranged from �200 ms to 800 ms.Statistical analyses were performed at each node in a 2D (channel � time

point) grid. For time-frequency analysis, preprocessed EEG data werelow-pass filtered (cutoff � 50 Hz). Event-related (log) spectral perturba-tion (ERSP) was calculated using Morlet wavelets (Delorme and Makeig,2004) at each frequency in theta (4 –7 Hz, 3 cycles), alpha (8 –13 Hz, 6cycles), and beta (14 –30 Hz, 10 cycles) bands. ERSP was computed at thetrial level and was then grouped and averaged based on experimentalconditions. The ERSP data spanned from �80 ms to 590 ms after theonset of the stimulus, with a sampling rate of 100 Hz. Statistical analyseswere performed at each of the nodes in a 3D (channel � time point �frequency) grid.

EEG data analysis. ERP and time-frequency data were divided intoconditions for each task phase. For the association phase, trials withinverted images were excluded. The remaining trials were divided into 16conditions, representing the 8 stimuli (4 tools and 4 landmarks) � 2 runbins (Runs 1 and 2 and Runs 3– 6). At the individual subject level, themean trial numbers were 257 (range: 208 –312) and 247 (range: 194 –314) for tools and landmarks, respectively. By contrasting the EEG sig-nals in the early (i.e., Runs 1 and 2) with the late part of the associationtask (i.e., Runs 3– 6; Fig. 1D), we investigated neural signals related to thegeneralization of associated CCD from tools to landmarks via the tool–landmark associations. The training phase data (219 trials on average,range: 150 –260) were partitioned into 4 conditions, representing the 2(associated CCD level: high vs low) � 2 (congruency) factorial design.CCD level was divided into 2 levels rather than a trial-level continuousvariable as in the model-based analysis. This approach was adopted outof concern that the signal-to-noise ratio in the EEG data at single chan-nels, time points, and frequency may be insufficient to ensure robustsignal in each node on each trial and hence may reduce statistical power.As shown in Figure 2A, model-estimated CCD shows a clear distinctionbetween tools with high and low CCD levels; thus, the 2 CCD levelsprovided a good approximation of the distribution of the model-derivedCCDs, and simultaneously enhanced sensitivity by averaging across mul-

Figure 2. Behavioral results. A–C, Training phase results. A, Individual model-estimated associated CCD levels, measuredas the mean Pi across all trials sharing the same CCD level, plotted as a function of CCD levels. Each line indicates one participant. B,Group mean�SEM of accuracy, plotted as a function of associated CCD and congruency. C, Individual estimated coefficients for theassociated CCD, congruency (Con), and control prediction error (CPE) regressors of the model-based analysis on RT. D, E, Groupmean � SEM of accuracy (D) and RT (E) in the test phase, plotted as a function of the CCD of the associated tool and congruency.


https://sccn.ucsd.edu/eeglab/index.phphttps://sccn.ucsd.edu/eeglab/index.php

tiple trials within each condition. The test phase data (mean trial number:144, range: 86 –167) were grouped based on a 2 (associated CCD ofpaired tool: high/low) � 2 (congruency/incongruency) factorial design.To avoid bias in ISPC, only landmarks with neutral ISPC (i.e., LL1 andLH1) were used in tests for generalization of CCD (mean trial number:72, range: 45– 84).

One main goal of this study is to test the temporal order of associatedand actual CCD effects. To avoid bias, we chose not to use predefinedtime windows for different effects, and instead adopted a data-drivenapproach that searched the whole temporal span of EEG data followingthe onset of the stimulus. Specifically, the statistical analyses of the effectsof interest were conducted using nonparametric cluster-based permuta-tion tests (Maris and Oostenveld, 2007). Dependent-sample t tests wereperformed to compare the conditions at every data node (electrode, timepoint, and frequency). Clusters of significant ( p � 0.01, one-tailed tests)adjacent nodes were identified and grouped together. Two nodes wereconsidered adjacent when they only differed in one dimension, with thedifference being within 4 cm Euclidean distance between channels, 1time point (10 ms), or 1 Hz. We then used the maxsum statistics, definedas the sum of the t statistics across all nodes within a cluster, as a summaryquantification of both the statistical significance within nodes and thespan of the cluster. The nonparametric cluster-based permutation testswere conducted separately for positive and negative statistics becausedifference in sign may indicate different neural processes. To determine pvalues that controlled for multiple comparisons, the maxsum of clusterswere compared with a null distribution, which was comprised of themaxsum of the clusters from 6000 simulations that repeated the sameanalysis with randomly shuffled condition labels (e.g., Bramão and Jo-hansson, 2017; Bramão et al., 2017). Due to the distinct neural processesrepresented by and the different number of cycles used for the theta,alpha, and beta bands, statistical analyses were conducted on these bandsseparately. The same threshold for statistical significance was applied toall frequency bands, allowing for comparison of temporal span acrossclusters from different frequency bands.

To test trial-level brain– behavior correlations, we extracted cluster-mean EEG data (e.g., theta power) for a given cluster from the nonpara-metric cluster-based permutation tests at each trial as a regressor, whichwas combined with a constant regressor to form a GLM. Then the trial-wise RT was regressed against this GLM to obtain the coefficient for theEEG data regressor, providing a quantification of the EEG data’s modu-lation on RT. Critically, because a significant CCD � congruency inter-action effect (i.e., the difference between when the [generalized]associated CCD matched the actual CCD and when they did not) wasused to identify the cluster for this analysis, we prevented double-dippingby performing this analysis separately for each condition of the CCD �congruency factorial design. The mean of the coefficients from the fourconditions was calculated for each participant and was passed on to agroup-level t test against 0 (i.e., no modulation of EEG data on behavior).

For two clusters found in the nonparametric cluster-based permuta-tion tests, we compared their temporal distributions marginalized overchannel and frequency (i.e., the likelihood of finding a data node in thecluster for a given time) to test which cluster emerged earlier. To this end,we formed the null hypothesis that variable A (representing a marginal-ized temporal distribution) with distribution PA precedes another ran-dom variable B with distribution PB. To test this hypothesis, wecalculated the probability that PA precedes a time point b randomly sam-pled from PB. Once b is drawn, we computed the probability of

P(A � b) ��0

bPA(a)da. Plugging PA � b into the random sampling

of b based on PB, we obtained the p value as the probability of the nullhypothesis being supported (i.e., PA � B), which takes the followingform:

P A � B � �PBb�0

b

PAadadb

In other words, PA � B denotes the probability of observing a � b byrandomly drawing a and b for infinite times. To avoid confusion with the

inferential statistics (see below), PA � B is henceforth referred to as“precedence index,” which ranges from 0 to 1. The lower the precedenceindex, the less likely A precedes B. In particular, a precedence index of 0.5indicates that A and B are equally likely to precede each other.

To account for sampling error, we estimated the distribution of theprecedence index using bootstrap resampling. Specifically, we randomlyresampled (with replacement) the subjects to form a new sample of 43subjects. We then performed the group-level nonparametric cluster-based permutation tests and identified the clusters showing highest max-sum statistics for each of the CCD effect and the CCD � congruencyinteraction effect. These two clusters were then submitted to the afore-mentioned temporal distribution comparison. This bootstrap resam-pling procedure was repeated for 1000 times and resulted in adistribution of the precedence index that the CCD � congruency inter-action effect preceded the CCD effect.

ResultsBehavioral dataParticipants performed the inversion detection task in the asso-ciation phase with high accuracy (group mean � SEM: 0.99 �0.004). On the rare trials in which participants needed to respondto inverted stimuli, the hit rate was 0.97 � 0.01. These resultsindicate that participants followed the task instructions and wereattentive to the images.

In the training phase, we first validated the model by compar-ing its prediction of Pi with the experimental manipulation oftool–CCD associations. Consistent with the task design, modelbelief of CCD for tools with high ISPC (0.68 � 0.01) was signif-icantly higher than for those with low ISPC (0.32 � 0.01, paired ttest: t(42) � 16.14, p � 0.001; Fig. 2A). Because the predictionsalso included the learning process, the model belief of ISPC isexpected to fall below the theoretical value (i.e., 0.25 and 0.75).Next, analysis of the effect of congruency on participant accuracyrevealed a significant main effect (F(1,42) � 36.43, p � 0.001),driven by higher accuracy on congruent (0.96 � 0.01) comparedwith incongruent trials (0.93 � 0.01). Neither the main effect ofISPC (F(1,42) � 0.20) nor the interaction between ISPC and con-gruency (F(1,42) � 1.13, p � 0.29) was significant (Fig. 2B). More-over, a model-based analysis on RT data replicated the classicfinding that incongruent trials were slower than congruent trials,evidenced by incongruency positively modulating RT (0.32 �0.02, t(42) � 11.67, p � 0.001; Fig. 2C, middle column).

More importantly, in the training phase, we expected that thepresentation of the tool would initiate retrieval of its associatedCCD, guiding conflict resolution. Thus, when the associatedCCD deviates from the actual CCD experienced as a function ofcongruency on the trial (i.e., when control prediction error islarge), retrieved CCD will mislead conflict resolution, resulting inslower responses (Jiang et al., 2014, 2015; Chiu et al., 2017;Muhle-Karbe et al., 2018). These predictions were confirmed by asignificant positive modulation of control prediction error on RT(0.32 � 0.12, t(42) � 2.66, p � 0.01; Fig. 2C, right column). Thisfinding indicates successful encoding of the item-specific CCD–tool associations and the influence of these associations in guid-ing cognitive control in the training phase. Finally, themodulation of associated CCD on RT was not significant(�0.11 � 0.28, t(42) � �0.38, p � 0.71). This null result is con-sistent with previous studies using ISPC manipulations (Chiu etal., 2017; Bejjani et al., 2018), and was expected based on afore-mentioned theory. This is because the difference between differ-ent levels of associated CCD is short-lived and will be replaced byactual CCD, leading to limited influence on the main effect ofassociated CCD in behavior. As a comparison, we also performeda repeated-measures 2 � 2 ANOVA on training phase RTs. There


was a significant main effect of congruency (F(1,42) � 151.13, p �0.001). Neither the main effect of associated CCD nor the inter-action was significant (both F values � 1; low ISPC/congruent:640 � 12 ms; low ISPC/incongruent: 680 � 13 ms; high ISPC/congruent: 640 � 11 ms; high ISPC/incongruent: 676 � 13 ms).Compared with the model-based analysis that specifically exam-ined the learning-related effect, the interaction effect may be con-founded by other factors, such as feature-binding (Mayr et al.,2003), perhaps contributing to this null result.

In the test phase, accuracy on landmarks with 50% ISPC (i.e.,LL1 and LH1) exhibited a significant main effect of congruency(F(1,42) � 15.59, p � 0.001; Fig. 2D), driven by higher accuracy oncongruent (0.95 � 0.01) than incongruent trials (0.91 � 0.01).Additionally, a marginally significant main effect of the ISPC ofthe associated tool (F(1,42) � 3.88, p � 0.06) evidenced a trend forhigher accuracy in the high CCD (LH1: 0.94 � 0.01) than lowCCD condition (LL1: 0.92 � 0.01). The interaction between thetwo factors was not significant (F(1,42) � 0.004). In RT data, thereagain was a significant main effect of congruency (F(1,42) � 98.75,p � 0.001; Fig. 2E), driven by faster responses in the congruent(670 � 11 ms) than incongruent condition (730 � 14 ms). Themain effect of the ISPC of the associated tool was not significant(F(1,42) � 1.07, p � 0.31).

Crucially, in the test phase, we observed a significant interac-tion between the CCD of the associated tool and congruency(F(1,42) � 9.82, p � 0.003). Similar to the control prediction erroreffect found during the training phase, the interaction exhibited apattern wherein mismatched CCD of the associated tool and ac-tual congruency (i.e., LL1 in incongruent trials and LH1 in con-gruent trials), which corresponded to larger prediction error, ledto slower RTs than matched conditions (i.e., LL1 in congruenttrials and LH1 in incongruent trials). Critically, the fact that thedirectly experienced ISPC in the test phase was the same for LL1and LH1 landmarks rules out the possibility that this interactionwas attributable to the test phase. Consistent with recent behav-ioral findings (Bejjani et al., 2018), this interaction effect indicatesthat the CCD linked to a tool was transferred to its associatedlandmark.

EEG results: validationAs the first step of EEG analysis, we validated the data by testingwhether they replicate the congruency effect (specifically, stron-ger mid-frontal negativity, sometimes followed by stronger pos-terior positivity on incongruent than congruent or neutral trials)found in previous ERP studies (Liotti et al., 2000; Folstein andVan Petten, 2008; Hanslmayr et al., 2008). Consistent with thesestudies, in the present experiment, ERP analyses revealed a maineffect of congruency in both the training (Fig. 3A) and test (Fig.3B) phases. Specifically, in the training phase, a cluster of midlineand frontal channels (Fig. 3C, leftmost panel, corrected p � 0.05)showed significantly greater positivity on congruent relative toon incongruent trials, starting �550 ms poststimulus onset (Fig.3C, second panel from left) and continuing until the end of thestimulus presentation (i.e., 800 ms). The ERP time courses acrossthese channels were similar in the test phase (Fig. 3C, right).Indeed, when testing the congruency effect in this cluster usingtest phase data, the pattern of greater positivity on congruentthan incongruent trials persisted (t(42) � 2.53, p � 0.008). Com-plementing the frontal, midline effect, a cluster of occipital chan-nels showed greater positivity on incongruent compared withcongruent trials (Fig. 3D, leftmost panel, corrected p � 0.05),diverging from �550 ms to �800 ms poststimulus onset in both

the training (Fig. 3D, second panel from left) and test phases (Fig.3D, second panel from right; t(42) � 3.10, p � 0.002).

EEG results: temporal dynamics of memory-guidedcognitive controlOur behavioral data revealed the involvement of associated CCDin guiding cognitive control (Fig. 2C). Based on the dual mecha-nisms of cognitive control theory (Braver et al., 2007; Braver,2012) and computational simulations (Blais et al., 2007; Vergutsand Notebaert, 2008), we hypothesized that, within a trial in thetraining and the test phases, cognitive control will first be guidedby the retrieved/predicted CCD and then gradually shift to theactual CCD (i.e., the experienced [in]congruency). Accordingly,in the EEG data, we expected that, within a trial, a main effect ofassociated CCD would be first observed, reflecting the retrieval ofthe associated CCD to guide cognitive control. Subsequently, ontrials in which the retrieved associated CCD conflicted with theactual CCD required to guide cognitive control, a mismatch ef-fect would signal the need and engagement in adjustment of con-trol. In other words, neural activity was expected to differbetween the scenario when the associated and actual CCDs wereconsistent (i.e., high CCD in incongruent trial and low CCD incongruent trial) and when they were inconsistent (i.e., low CCDin incongruent trial and high CCD in congruent trial), leading toan interaction effect between associated CCD and congruency(Fig. 4). We did not consider the main effect of congruency be-cause it may reflect reactive cognitive control (i.e., withholdingadjustment of cognitive control until the detection of actualCCD), rather than the proactive cognitive control focused on inthis study.

To test these predictions, we performed 2 (associated CCD:high/low) � 2 (congruency/incongruency) repeated-measuresANOVAs on the ERP data and the time-frequency signals (theta,alpha, and beta bands) from the training phase. Multiple com-parisons were corrected for using nonparametric cluster-basedpermutation tests. To test the neural processes shared by bothtraining and test phase data (e.g., the generalization of the asso-ciated CCD) and to examine the validity of the findings, we usedclusters detected in the training phase as ROIs and repeated theanalyses using the ROIs in the test phase data. To be consistentwith the hypothetical chronological order shown in Figure 4, wefirst present the results of the main effect of associated CCD andthen the results of the associated CCD � congruency interaction.

Early in the training phase (peaking at �200 ms), we observedlower alpha-band ERSP on high associated CCD trials than onlow associated CCD trials (Fig. 5A). A nonparametric cluster-based permutation test statistically confirmed that the high CCDcondition was associated with lowered ERSP in the alpha-band ina group of left frontal and middle channels (corrected p � 0.048;Fig. 5C, left) and peaked at �200 ms after stimulus onset (Fig. 5C,middle, right). No other clusters passed the nonparametriccluster-based permutation tests.

We next turned to the test phase data and used this cluster totest the generalization of the CCD to landmarks. A similar spa-tiotemporal pattern was found in the test phase when comparinghigh and low transferred CCD trials (i.e., LH1 vs LL1; Fig. 5B).Consistent with the training phase data, we observed statisticallysignificant lower alpha-band ERSP for the high transferred CCDlandmark (i.e., LH1) compared with low transferred CCD land-mark (i.e., LL1, t(42) � 2.13, p � 0.04; Fig. 5D, left). In the chan-nels showing an associated CCD effect in the training phase(Fig. 5C, left), the generalized CCD effect at test demonstratedtemporal and frequency spans similar to those in the training


phase (Fig. 5D, middle, right). As such, the (generalized) CCDeffects presently observed in two independent sets of EEG data(i.e., training and test phases) strongly support the notion thatthis cluster represents engagement of the (generalized) asso-ciated CCD.

Later within trials during the trainingphase, we observed an interaction be-tween associated CCD � congruency inthe theta-band ERSP (Fig. 6A), revealing acluster that included posterior midlinechannels that peaked at �350 ms afterstimulus onset (Fig. 6C, left; corrected p �0.044). Numerically, the test phase datadisplayed a similar interaction in the pos-terior midline channels (Fig. 6B), whichreflected lower ERSP for trials with mis-matched associated CCD and congruency(i.e., incongruent trials with low associ-ated CCD and congruent trials with highassociated CCD) compared with trialswith matched associated CCD and con-gruency (i.e., incongruent trials with high

associated CCD and congruent trials with low associated CCD;Fig. 6C, middle, right). We further found that, at the trial level,theta-band power in the cluster in Figure 6C explained variancein RT, such that lower theta power (reflecting mismatch between

Figure 3. Replication of congruency effect in ERP data. A, B, Spatiotemporal distribution of congruency effect in the training and test phases, respectively. C, D, From left to right: Topographicmaps of the duration over which the congruency effect was significant in each channel of the cluster identified by the nonparametric cluster-based permutation tests; temporal distribution ofsignificant congruency effects in nodes in the cluster; ERP time courses of representative channels, plotted by experimental conditions, in the training and test phases, respectively. Con, Congruenttrials; Inc, incongruent trials. For all time course plots, the temporal resolution is 2 ms.

Figure 4. Hypothetical time course of processes driving the engagement of cognitive control (top) and of the underlying EEGeffects (bottom).


associated CCD and congruency, or larger control predictionerror) was accompanied by slower responses (collapsed acrossconditions in the CCD � congruency factorial design to decon-found the shared variance in the CCD � congruency interactionfound in both behavioral and EEG data, t(42) � �3.31, p � 0.002;Fig. 6D). When applying this cluster to test phase data, the inter-action effect, which was defined following Figure 6A to preservethe sign of the effect, was also significant (t(42) � �2.13, p � 0.04;Fig. 6E), and the interaction effect also predicted test phase RT atthe trial level (t(42) � �2.18, p � 0.03; Fig. 6F). The behavioralrelevance of this cluster suggests that it is involved in the onlineadjustment of cognitive control that shifts from memory-guidedto actual CCD-guided control.

In the training phase, one other posterior midline cluster dis-played a significant CCD � congruency interaction effect in thealpha-band; this effect peaked at �300 ms after stimulus onset(corrected p � 0.006; compare Fig. 6C, middle). However, whenreplicating the aforementioned brain– behavior analysis, alpha-

band power in this cluster did not significantly explain variancein RT in training phase data (t(42) � �1.79, p � 0.08). Further-more, when we repeated these analyses using the test phasedata, neither the interaction effect (t(42) � �1.51, p � 0.14)nor the brain– behavior analysis was significant (t(42) � �0.12,p � 0.9). Due to the lack of replicability and behavioral rele-vance, we conclude that activity in this cluster does not reflectthe adjustment of cognitive control following the detection ofthe actual CCD. No other clusters passed the nonparametriccluster-based permutation tests.

ERP analyses revealed neither a main effect of associated CCDnor an associated � actual CCD interaction that survived thenonparametric cluster-based permutation tests. These null re-sults may reflect phase incoherence in event-induced activityacross trials (Bastiaansen and Hagoort, 2003). When repeatingthe analyses of main effect of associated CCD and CCD � con-gruency interaction using test phase data, the nonparametriccluster-based permutation tests did not reveal any significant re-

Figure 5. Alpha-band oscillations show early (generalized) CCD effects in left middle and frontal channels. A, B, Spatiotemporal distribution of alpha-band CCD effect (high vs low) in the trainingand test phases, respectively. C, A cluster showing the associated CCD effect in the training phase. Left, Visualization of channel-wise proportion of significant observations (each observation isdefined as a combination of channel, time point, and frequency) in the cluster showing a significant associated CCD effect. Middle, Size of the associated CCD effect in the cluster, plotted as a functionof frequency and time. The cluster is highlighted by the red box. Right, Time course of cluster-mean (� SEM) ERSP change relative to time course grand mean, plotted as a function of associated CCDlevels. D, Leftmost panel, Test phase individual ERSP averaged within the cluster plotted as a function of generalized CCD levels (left) and group mean (� SEM) ERSP difference between LL1 and LH1(right). The other two panels are organized similar to C.


sults. This was possibly due to the low trial count and subsequentlow signal-to-noise ratio at the node level.

None of the three frequency bands showed a significant effectof congruency. Given that the time window of the time-frequencyanalyses ended at 590 ms after stimulus onset, the lack of a sig-nificant effect here does not contradict the ERP findings showinga congruency effect after 550 ms after stimulus onset. We specu-late that the relatively late congruency effects reflect the domi-nance of actual CCD, after correcting the associative CCD, inguiding cognitive control (reflected in the associated � actualCCD interaction).

To determine whether proactive control adjustments in re-sponse to anticipated CCD precedes further control adjustmentsin response to actual CCD, we quantitatively tested whetherthe associated CCD effect temporally preceded the CCD �congruency interaction. For each of the clusters showing the

associated CCD effect (Fig. 5C) and the CCD � congruencyinteraction (Fig. 6C), we calculated their marginalized proba-bilistic density functions on the temporal dimension (Fig. 7)and calculated the precedence index that the distribution ofthe associated CCD effect followed the distribution of the in-teraction effect (see Materials and Methods). A resulting pre-cedence index of 0.02 suggested a high chance that the CCD �congruency interaction effect occurred after the CCD effect.We also estimated the distribution of the precedence index bypercentile bootstrapping 1000 times (see Materials and Meth-ods). The nonparametric 95% CI of the precedence index was[0.0048, 0.4464], which lay outside of the baseline value of 0.5.This result indicates a p value �0.05 for the null hypothesisthat the CCD effect did not precede the CCD � congruencyinteraction effect.

Figure 6. Theta-band oscillations show interaction between associated CCD and congruency in posterior midline channels. A, B, Spatiotemporal distribution of theta-band associated CCD �congruency interaction in training phase (A) and test phase (B). C, A posterior midline cluster showing significant theta-band associated CCD � congruency interaction. From left to right:Visualization of channel-wise proportion of significant observations in the cluster; size of interaction effect with the cluster highlighted in red box, plotted as a function of frequency and time; timecourse of cluster-mean ERSP, plotted as a function of associated CCD levels. D, Group mean RT (� SEM) in the training phase, plotted as a function of quintiles of the theta-band associated CCD �congruency interaction effect. RTs in the test phase are visualized similarly in F. E, ERSP when applied to the cluster in C on test phase data, plotted as a function of congruency and (transferred)associated CCD level.


EEG results: generalization of CCD throughtool–landmark associationAs reported above, even when landmarks LL1 and LH1 werepresented with the same neutral ISPC (50% congruency) duringthe test phase, behavioral analyses of RT revealed a significantinteraction between the indirectly paired CCD (i.e., through thelandmark’s associated tool) and congruency (Fig. 2E), and neuralanalyses revealed a main effect of the indirectly paired CCD onalpha-band oscillations (Fig. 5D). These results provide strongevidence that participants generalized the learned CCD fromtools to their associated landmarks.

In a final set of analyses, we explored the possible mechanismssupporting this generalization. In particular, we hypothesizedthat, during the interleaved presentations of tool–landmark pairsin the association phase and tool–CCD pairs in the trainingphase, these two associations become integrated in memory,forming a tool–landmark–CCD representation that enables thegeneralization of CCD to the landmark (Fig. 8A).

One potential task phase during which integrative encodingmay occur is in Runs 3– 6 of the association phase (i.e., after initialexposure and encoding of the tool–CCD associations; Fig. 1D).During the time course of a trial in these runs, presentation of thetool may reactivate the learned tool–CCD association, which canthen be associated with the subsequent landmark upon its pre-sentation. To test this possibility, we compared EEG data follow-ing the presentation of landmark images to that following thepresentation of tool images in Runs 3– 6. The tool image datawere used as a baseline to tease apart EEG data reflecting nuisanceprocesses, such as perceptual processing. Data from the first 2runs of the association phase were included as an additional baselinewhen tool–CCD associations were not available. This baselinefilters out neural activity of the encoding of the tool–landmarkassociation and helps isolate signal potentially reflecting integra-tive encoding. Therefore, the test for integrative encoding tookthe form of an interaction between stimulus category (tool vslandmark) and time (Runs 1 and 2 vs Runs 3– 6), as shown inFigure 8B.

Across the ERP data and the time-frequency data, the non-parametric cluster-based permutation tests revealed an alpha-band, mediofrontal cluster centered at �300 ms after onset of thelandmark image (corrected p � 0.036; Fig. 8C,D, left, middle).This effect was driven by data from Runs 3– 6, which showedreduced alpha-band ERSP following the onset of the landmarks,compared with following the presentation of the tools (Fig. 8D,right). Given that memory retrieval is accompanied by alpha de-

synchronization (Hanslmayr et al., 2012, 2016), the increasedalpha-band ERSP in tools than landmarks was unlikely to reflectthe retrieval of the tool–CCD association. Critically, to testwhether this interaction effect was linked to the generalization ofCCD to the landmarks, we performed a cross-participant corre-lation analysis between the cluster-average interaction effect foreach participant and the cluster-mean of the transferred CCDeffect in the test phase (Fig. 5D) for that participant. Similar to thebehavioral analysis, this analysis was performed on items withneutral test phase ISPC (i.e., LL1 and LH1), to deconfound theISPC in the test phase. Results revealed a significant positive re-lationship (r � 0.38, p � 0.01; Fig. 8E, top), indicating that par-ticipants with stronger post-landmark alpha power decrease inthe association phase tended to show a stronger alpha-bandtransferred CCD effect in the test phase. To examine whether thiseffect was item-specific (i.e., only occurring within the sameitems), we repeated this analysis by keeping test phase data un-changed while replacing association phase items LL1 and LH1with different items LL2 and LH2, thus forming a cross-itemdesign. A positive correlation coefficient would be evidenceagainst the item-specific claim. However, a negative correlation(r � �0.35, p � 0.02; Fig. 8E, bottom) was observed, thus sup-porting the item-specific claim. The negative correlation showsthat post-landmark alpha power decreased in the associationphase between LL1-LH1 and LL2-LH2 (r � �0.71, p � 0.001),which may reflect interference between individual associationsduring generalization.

An alternative interpretation of the stimulus category � timeinteraction may be that it reflects differential processing of orattention to the tool images relative to the landmark images.Specifically, because each landmark appears 15 times in each as-sociation phase run and each tool appears 15 times in each asso-ciation phase run and 12 times in each training phase run (whichwere interleaved with association phase runs), a change in theEEG response to a stimulus between association phase Runs 1and 2 and Runs 3– 6 can be viewed as an adaptation effect. Fromthis perspective, the interaction effect might reflect strongeradaption effects for tools than landmarks, given that tools wereencountered more often (due to their presentation in the inter-leaved training phase runs). At the individual level, stronger ad-aptation for tools than landmarks (compare Fig. 8D, right) mightbe attributed to more attention to the tool images in the trainingphase; such attention should impact tool processing during thetraining phase and thus impact the magnitude of the congruencyeffect (e.g., attention-enhanced processing of the task-relevantstimulus [i.e., the tool image] might reduce the congruency ef-fect). In short, this alternative interpretation predicts that a stron-ger stimulus category � time interaction in the association phasewill be related to a weaker congruency effect in the training phase.To test this prediction, we used cross-subject correlational anal-yses between the stimulus category � time interaction effect inthe association phase and the congruency effect in the trainingphase. Given that significant training phase congruency effectswere observed in the ERP (clusters identified in Fig. 3C,D), be-havioral accuracy, and RT data (Fig. 2), we conducted four anal-yses, each using one effect as a measure of the congruency effect.None of the correlations reached significance (all p values �0.18). Therefore, the stimulus category � time interaction effectappears less likely to reflect differential adaptation to the toolsthan the landmarks.

Another possibility is that the cluster reflected the change inthe predictability/association strength between tool and land-mark. If this were true, we would predict that this effect (reflect-

Figure 7. Temporal distribution of the clusters showing main effect of associated CCD andassociated CCD � congruency interaction.


ing tool–landmark association) and the main effect of associatedCCD in the training phase (reflecting tool–CCD association; Fig.5C) will jointly predict the main effect of generalized CCD in thetest phase EEG data (Fig. 5D). To test this prediction, we con-ducted cross-subject rescaling (range: 0 –1) separately on the ef-fect of tool minus landmark in the cluster reported in Figure 8Dand the training phase associated CCD effect. These rescaled ef-fects for TL1, TH1, LL1, and LH1 were combined and correlatedwith the main effect of generalized CCD for LL1 and LH1 in thetest phase. For the joint prediction, we tested whether thestrength of generalized CCD relates to either (1) the sum ofthe two predictor effects or (2) the product of the two predictoreffects. Neither yielded a significant correlation with the general-

ized CCD effect (both p values � 0.18). Thus, it is unlikely thatthis effect reflects the predictability/association strength betweenthe tool and the landmark.

DiscussionHow cognitive control is regulated is of key interest in under-standing goal-directed behavior (Botvinick et al., 2001; Botvinickand Cohen, 2014; Waskom et al., 2017). Recent theoretic ad-vances propose that cognitive control can be proactively adjustedbased on prediction of future CCD (Botvinick et al., 2001; Brownand Braver, 2005; Braver et al., 2007; Braver, 2012). A wealth ofdata indicate that such predictions can be based on temporalinformation (e.g., Logan and Zbrodoff, 1979; Botvinick et al.,

Figure 8. Time-frequency analysis and results in the association phase. A, The integrative encoding hypothesis. B, The contrast for the stimulus category � time interaction. C, Spatiotemporaldistribution of the interaction effect following the onset of the stimulus. D, A mediofrontal cluster showing significant alpha-band stimulus category � time interaction. From left to right:Visualization of channel-wise weights in the cluster; size of interaction effect plotted as a function of frequency and time with the cluster highlighted in red box; time course of cluster-mean (� SEM)ERSP plotted as a function of experimental conditions. E, Transferred CCD effect in the test phase, measured by the alpha-band ERSP difference between LL1 and LH1 in the cluster shown in Figure5C, plotted out as a function of the stimulus category � time interaction effect for the same items (left) and different items (right) in the association phase.


1999; Carter et al., 2000; Kerns et al., 2004; Egner and Hirsch,2005; Egner, 2007; Hazeltine et al., 2011; Aben et al., 2019; DeLoof et al., 2019). Here, we explored another important source ofCCD predictions: learned associations between items and CCD(e.g., Jacoby et al., 2003; Bugg et al., 2011; Chiu et al., 2017). Toadvance understanding of (1) how item-CCD associative mem-ories proactively guide the regulation of cognitive control and (2)the neurocognitive mechanisms supporting the generalization ofCCD, we leveraged behavioral and EEG measures acquired whileparticipants performed a three-phase task that involved thelearning of tool–landmark and tool–CCD associations and theassessment of the generalization of CCD from tools to landmarks(Fig. 1). Our findings provide novel evidence for memory-guidedproactive adjustments of control that precede adjustments due toactual demands, and the generalization of CCD via associativememory.

Temporal dynamics of memory-guided adjustments ofcognitive controlWe first applied a reinforcement learning model (Chiu et al.,2017) to the behavioral data to quantify how associated CCD andactual CCD jointly affect behavior. Analyses revealed that train-ing phase RT scaled with the degree of discrepancy between as-sociated and actual CCD. Based on the theory of proactivecognitive control (Braver et al., 2007; Braver, 2012) and compu-tational simulations (Blais et al., 2007; Verguts and Notebaert,2008), this observation suggests that the retrieval of associatedCCD leads to proactive adjustments of control that speed goal-directed behavior when predicted and actual CCD align; by con-trast, when misaligned, additional reactive adjustments arerequired based on actual CCD following its detection. Consistentwith this interpretation, analyses of the EEG data indicated thatthe effect of associated CCD emerged earlier than the effect ofactual CCD. Specifically, lower alpha-band ERSP, at left channelsand peaking at �200 ms after stimulus onset, was found in trialswith higher associated CCD (Fig. 5C). This decrease in alphaoscillations may reflect the increased involvement of selectiveattention (Klimesch, 1999; Sadaghiani and Kleinschmidt, 2016)in anticipation of the forthcoming incongruent trial predicted bythe higher associated CCD. This converges with fMRI data dem-onstrating that higher associative CCD was accompanied bygreater activation in dorsolateral PFC and ACC (Blais and Bunge,2010).

Subsequent to the associated CCD effect (Fig. 7), we observedan interaction between associated and actual CCD in theta-bandERSP in posterior channels (Fig. 6C). This interaction effect isconsistent with similar interactions in ERP when participantsperform the Stroop (Shedden et al., 2013) and Simon (Whiteheadet al., 2017) tasks. In our data, this interaction appears to bedriven by increased theta-band oscillations when the associatedCCD matched the actual CCD (Fig. 6C, right). While speculative,compared with a mismatch that requires further adjustment ofcognitive control, a match may lead to elevated readiness of in-formation processing, which has been associated with highertheta oscillation (Basar et al., 2001). Crucially, the theta-bandERSP in these channels was negatively correlated with RT at thetrial level (Fig. 6D). This result is consistent with previous find-ings that increased task-elicited theta is accompanied by betterperformance (Klimesch, 1999), and suggests that this interactionrelates to resolution of the discrepancy between associated andactual CCD (as slower RTs indicate larger discrepancies that mustbe resolved). Importantly, we cross-validated these findings, ob-

serving similar results using the independent data from the testphase (Figs. 5B,D, 6B, 8B,D).

Generalization of CCD through item-item associationsOur behavioral and EEG data provide strong evidence of transferof CCD from tool images to their associated landmarks. Behav-iorally, we found an interaction between the CCD of the associ-ated tool and congruency in LL1 and LH1 trials (Fig. 2E) in thetest phase. This finding replicates recent work (Bejjani et al.,2018). In the EEG data, we observed a significant main effect ofthe associated tool’s CCD on landmark-triggered alpha oscilla-tions during the test phase (Fig. 5D). Critically, neither findingcan be attributed to test-phase learning processes, given that theISPCs for LL1 and LH1 items were identical in the test phase. Theonly difference between LL1 and LH1 items was the level of CCDpreviously bound to their associated tools; as such, these behav-ioral and EEG differences between these landmarks documentthe generalization of CCD from tools to landmarks.

Regarding the neurocognitive mechanisms of generalization,one possibility is that generalization occurred through integrativeencoding (Shohamy and Wagner, 2008) that merged tool–CCDassociations and tool–landmark associations into conjunctivetool–landmark–CCD memories (Fig. 8A). Supporting this idea,we observed differential alpha-band ERSP in medial frontalchannels following repeated presentation of landmarks and toolsin the association phase (tool/landmark � Runs 1 and 2/Runs3– 6; Fig. 8D). This decrease may be linked to generalization be-cause it emerged following exposure to the tool–CCD associa-tions in the training phase (Fig. 8D), and it predicted themagnitude of the generalized CCD effect in the test phase (Fig.8E). Prior work has linked decreases in alpha oscillations to mem-ory encoding and retrieval (Hanslmayr et al., 2012), and havebeen posited to reflect desynchronization in cortical activationthat signals a shift from processing present inputs to memoryoperations (Hanslmayr et al., 2012, 2016). Moreover, prior ob-servations indicate that hippocampal processes support memorygeneralization (Shohamy and Wagner, 2008; Zeithamova andPreston, 2010; Kuhl et al., 2011; Wimmer and Shohamy, 2012;Zeithamova et al., 2012). Thus, our findings provided new in-sights about the electrophysiological mechanisms in the general-ization of abstract concepts, such as CCD, through partiallyoverlapping memories.

As an additional but nonexclusive mechanism, integrative en-coding may also occur in the training phase of the present para-digm. Specifically, as the tool–CCD pairings were experienced,presentation of the tool may have triggered retrieval of the asso-ciated landmark, providing an opportunity for forming an inte-grative memory. The present experimental design is not suitablefor examining whether integrative encoding also occurs in thetraining phase because this phase lacks baseline conditions thatare required to deconfound nuisance effects (e.g., there were notraining runs that occurred before initial exposure of tool–land-mark associations). We also did not find a direct correlation be-tween the sizes of training phase-associated CCD effect and thoseof the test phase-transferred CCD effect across participants (r ��0.19, p � 0.21). Future studies can examine this hypothesis bymoving the first training runs before the first association runs.Alternatively, generalization during the training phase could betested by examining neural evidence for reinstatement of the to-be-generalized item (Wimmer and Shohamy, 2012; Kurth-Nelson et al., 2015).

Another possibility is that generalization of CCD occurs atretrieval (i.e., the test phase) through inference over partially


overlapping associations, which requires additional time to se-quentially activate multiple overlapping associations for general-ization to occur (Kumaran and McClelland, 2012; Horner et al.,2015; Koster et al., 2018). Whereas Shohamy and Wagner (2008),among others, provided evidence that integration can occur dur-ing learning and before the critical generalization test, a recentstudy reported slower responses when participants made judg-ments based on retrieval of direct associations relative to thosebased on inferred associations (Koster et al., 2018). In the presentstudy, our finding that the associated CCD effect preceded actualCCD effects in guiding cognitive control suggests that generaliza-tion occurred via integration. Furthermore, the neural timing ofthe observed CCD effect in the training phase was similar to thatof the generalized CCD effect in test phase (Fig. 5A,C,D). Thisresult suggests a direct retrieval of the already generalized land-mark–CCD association, thus favoring an integrative encodingaccount.

A potential confound in the experimental design is that thegeneralization effect may co-occur with other processes that alsochange over time. Although we ruled out two confounds (adap-tation and tool–landmark association strength), future studiesshould explore the possible influence of other processes that varyover time.

In conclusion, this study provided new insights into the mech-anisms of associative memory-guided adjustment of cognitivecontrol. Specifically, supporting the hypothesis of an earlier in-volvement of associated CCD than actual CCD in guiding cogni-tive control, we found an early-onset–associated CCD effect inalpha oscillations. This effect was temporally followed by an in-teraction between associative and actual CCD in theta oscilla-tions, possibly reflecting their competition in guiding cognitivecontrol. Furthermore, supporting an integrative encoding ac-count, a generalized associated CCD effect in alpha oscillations inthe test phase was linked to a decrease in alpha oscillation duringthe encoding of item–item associations. These findings advanceunderstanding of the neurocognitive mechanisms supportingmemory-guided cognitive control during goal-directed behavior.

ReferencesAben B, Calderon CB, Van der Cruyssen L, Picksak D, Van den Bussche E,

Verguts T (2019) Context-dependent modulation of cognitive controlinvolves different temporal profiles of fronto-parietal activity. Neuroim-age 189:755–762.

Abrahamse E, Braem S, Notebaert W, Verguts T (2016) Grounding cogni-tive control in associative learning. Psychol Bull 142:693–728.

Basar E, Schürmann M, Sakowitz O (2001) The selectively distributed thetasystem: functions. Int J Psychophysiol 39:197–212.

Bastiaansen M, Hagoort P (2003) Event-induced theta responses as a windowon the dynamics of memory. Cortex 39(4 –5):967–972.

Bejjani C, Zhang Z, Egner T (2018) Control by association: transfer of im-plicitly primed attentional states across linked stimuli. Psychon Bull Rev25:617– 626.

Blais C, Bunge S (2010) Behavioral and neural evidence for item-specificperformance monitoring. J Cogn Neurosci 22:2758 –2767.

Blais C, Robidoux S, Risko EF, Besner D (2007) Item-specific adaptationand the conflict-monitoring hypothesis: a computational model. PsycholRev 114:1076 –1086.

Botvinick MM, Cohen JD (2014) The computational and neural basis ofcognitive control: charted territory and new frontiers. Cogn Sci 38:1249 –1285.

Botvinick MM, Braver TS, Barch DM, Carter CS, Cohen JD (2001) Conflictmonitoring and cognitive control. Psychol Rev 108:624 – 652.

Botvinick M, Nystrom LE, Fissell K, Carter CS, Cohen JD (1999) Conflictmonitoring versus selection-for-action in anterior cingulate cortex. Na-ture 402:179 –181.

Braem S, Bugg JM, Schmidt JR, Crump MJC, Weissman DH, Notebaert W,

Egner T (2019) Measuring adaptive control in conflict tasks. TrendsCogn Sci 23:769 –783.

Bramão I, Johansson M (2017) Benefits and costs of context reinstatementin episodic memory: an ERP study. J Cogn Neurosci 29:52– 64.

Bramão I, Karlsson A, Johansson M (2017) Mental reinstatement of encod-ing context improves episodic remembering. Cortex 94:15–26.

Braver TS (2012) The variable nature of cognitive control: a dual mecha-nisms framework. Trends Cogn Sci 16:106 –113.

Braver TS, Gray JR, Burgess GC (2007) Explaining the many varieties ofworking memory variation: dual mechanisms of cognitive control. In:Variation in working memory (Conway A, Jarrold C, Kane M, Miyake A,Towse J, eds), pp 76 –106. Oxford: Oxford UP.

Brown JW, Braver TS (2005) Learned predictions of error likelihood in theanterior cingulate cortex. Science 307:1118 –1121.

Bugg JM, Jacoby LL, Chanani S (2011) Why it is too early to lose control inaccounts of item-specific proportion congruency effects. J Exp PsycholHum Percept Perform 37:844 – 859.

Carter CS, Macdonald AM, Botvinick M, Ross LL, Stenger VA, Noll D, CohenJD (2000) Parsing executive processes: strategic vs evaluative functionsof the anterior cingulate cortex. Proc Natl Acad Sci U S A 97:1944 –1948.

Chiu YC, Egner T (2019) Cortical and subcortical contributions to context-control learning. Neurosci Biobehav Rev 99:33– 41.

Chiu YC, Jiang J, Egner T (2017) The caudate nucleus mediates learning ofstimulus-control state associations. J Neurosci 37:1028 –1038.

Cohen JD, Dunbar K, McClelland JL (1990) On the control of automaticprocesses: a parallel distributed processing account of the Stroop effect.Psychol Rev 97:332–361.

Crump MJ, Milliken B (2009) The flexibility of context-specific control:evidence for context-driven generalization of item-specific control set-tings. Q J Exp Psychol 62:1523–1532.

Delorme A, Makeig S (2004) EEGLAB: an open source toolbox for analysis ofsingle-trial EEG dynamics including independent component analysis.Journal of neuroscience methods 134:9 –21.

De Loof E, Vassena E, Janssens C, De Taeye L, Meurs A, Van Roost D, Boon P,Raedt R, Verguts T (2019) Preparing for hard times: scalp and intracra-nial physiological signatures of proactive cognitive control. Psychophys-iology 56:e13417.

Egner T (2007) Congruency sequence effects and cognitive control. CognAffect Behav Neurosci 7:380 –390.

Egner T (2014) Creatures of habit (and control): a multi-level learning per-spective on the modulation of congruency effects. Front Psychol 5:1247.

Egner T (2017) The Wiley handbook of cognitive control. Chichester, UK:Wiley Blackwell.

Egner T, Hirsch J (2005) Cognitive control mechanisms resolve conflictthrough cortical amplification of task-relevant information. Nat Neurosci8:1784 –1790.

Folstein JR, Van Petten C (2008) Influence of cognitive control and mis-match on the N2 component of the ERP: a review. Psychophysiology45:152–170.

Hanslmayr S, Pastötter B, Bäuml KH, Gruber S, Wimber M, Klimesch W(2008) The electrophysiological dynamics of interference during theStroop task. J Cogn Neurosci 20:215–225.

Hanslmayr S, Staudigl T, Fellner MC (2012) Oscillatory power decreasesand long-term memory: the information via desynchronization hypoth-esis. Front Hum Neurosci 6:74.

Hanslmayr S, Staresina BP, Bowman H (2016) Oscillations and episodicmemory: addressing the synchronization/desynchronization conun-drum. Trends Neurosci 39:16 –25.

Hazeltine E, Lightman E, Schwarb H, Schumacher EH (2011) The bound-aries of sequential modulations: evidence for set-level control. J Exp Psy-chol Hum Percept Perform 37:1898 –1914.

Horner AJ, Bisby JA, Bush D, Lin WJ, Burgess N (2015) Evidence for holisticepisodic recollection via hippocampal pattern completion. Nat Commun6:7462.

Jacoby LL, Lindsay DS, Hessels S (2003) Item-specific control of automaticprocesses: Stroop process dissociations. Psychon Bull Rev 10:638 – 644.

Jiang J, Heller K, Egner T (2014) Bayesian modeling of flexible cognitivecontrol. Neurosci Biobehav Rev 46:30 – 43.

Jiang J, Beck J, Heller K, Egner T (2015) An insula-frontostriatal networkmediates flexible cognitive control by adaptively predicting changingcontrol demands. Nat Commun 6:8165.

Kerns JG, Cohen JD, MacDonald AW 3rd, Cho RY, Stenger VA, Carter CS


(2004) Anterior cingulate conflict monitoring and adjustments in con-trol. Science 303:1023–1026.

King JA, Korb FM, Egner T (2012) Priming of control: implicit contextualcuing of top-down attentional set. J Neurosci 32:8192– 8200.

Klimesch W (1999) EEG alpha and theta oscillations reflect cognitive andmemory performance: a review and analysis. Brain Res Brain Res Rev29:169 –195.

Koster R, Chadwick MJ, Chen Y, Berron D, Banino A, Düzel E, Hassabis D,Kumaran D (2018) Big-loop recurrence within the hippocampal systemsupports integration of information across episodes. Neuron 99:1342–1354.e6.

Kuhl BA, Shah AT, DuBrow S, Wagner AD (2010) Resistance to forgettingassociated with hippocampus-mediated reactivation during new learn-ing. Nat Neurosci 13:501–506.

Kuhl BA, Rissman J, Chun MM, Wagner AD (2011) Fidelity of neural reac-tivation reveals competition between memories. Proc Natl Acad Sci U S A108:5903–5908.

Kumaran D, McClelland JL (2012) Generalization through the recurrentinteraction of episodic memories: a model of the hippocampal system.Psychol Rev 119:573– 616.

Kumaran D, Hassabis D, McClelland JL (2016) What learning systems dointelligent agents need? Complementary learning systems theory up-dated. Trends Cogn Sci 20:512–534.

Kurth-Nelson Z, Barnes G, Sejdinovic D, Dolan R, Dayan P (2015) Tempo-ral structure in associative retrieval. eLife 4:04919.

Leys C, Ley C, Klein O, Bernard P, Licata L (2013) Detecting outliers: do notuse standard deviation around the mean, use absolute deviation aroundthe median. J Exp Soc Psychol 49:764 –766.

Liotti M, Woldorff MG, Perez R, Mayberg HS (2000) An ERP study of thetemporal course of the Stroop color-word interference effect. Neuropsy-chologia 38:701–711.

Logan GD, Zbrodoff NJ (1979) When it helps to be misled: facilitative ef-fects of increasing the frequency of conflicting stimuli in a Stroop-liketask. Mem Cogn 7:166 –174.

Maris E, Oostenveld R (2007) Nonparametric statistical testing of EEG- andMEG-data. J Neurosci Methods 164:177–190.

Mayr U, Awh E, Laurey P (2003) Conflict adaptation effects in the absence ofexecutive control. Nature neuroscience 6:450 – 452.

Miller EK, Cohen JD (2001) An integrative theory of prefrontal cortex func-tion. Annu Rev Neurosci 24:167–202.

Muhle-Karbe PS, Jiang J, Egner T (2018) Causal evidence for learning-dependent frontal lobe contributions to cognitive control. J Neurosci38:962–973.

Poldrack RA, Packard MG (2003) Competition among multiple memorysystems: converging evidence from animal and human brain studies.Neuropsychologia 41:245–251.

Sadaghiani S, Kleinschmidt A (2016) Brain networks and alpha-oscillations:structural and functional foundations of cognitive control. Trends CognSci 20:805– 817.

Shedden JM, Milliken B, Watter S, Monteiro S (2013) Event-related poten-tials as brain correlates of item specific proportion congruent effects.Conscious Cogn 22:1442–1455.

Shohamy D, Wagner AD (2008) Integrating memories in the human brain:hippocampal-midbrain encoding of overlapping events. Neuron 60:378 –389.

Surrey C, Dreisbach G, Fischer R (2017) Context-specific adjustment ofcognitive control: transfer of adaptive control sets. Q J Exp Psychol(Hove) 70:2386 –2401.

Sutton RS, Barto AG (2018) Reinforcement learning: an introduction, Ed 2.Cambridge, MA: Massachusetts Institute of Technology.

Verguts T, Notebaert W (2008) Hebbian learning of cognitive control: deal-ing with specific and nonspecific adaptation. Psychol Rev 115:518 –525.

Waskom ML, Kumaran D, Gordon AM, Rissman J, Wagner AD (2014)Frontoparietal representations of task context support the flexible controlof goal-directed cognition. J Neurosci 34:10743–10755.

Waskom ML, Frank MC, Wagner AD (2017) Adaptive engagement of cog-nitive control in context-dependent decision making. Cereb Cortex27:1270 –1284.

Weidler BJ, Bugg JM (2016) Transfer of location-specific control to un-trained locations. Q J Exp Psychol (Hove) 69:2202–2217.

Whitehead PS, Brewer GA, Blais C (2017) ERP evidence for conflict in con-tingency learning. Psychophysiology 54:1031–1039.

Wimmer GE, Shohamy D (2012) Preference by association: how memorymechanisms in the hippocampus bias decisions. Science 338:270 –273.

Zeithamova D, Preston AR (2010) Flexible memories: differential roles formedial temporal lobe and prefrontal cortex in cross-episode binding.J Neurosci 30:14676 –14684.

Zeithamova D, Dominick AL, Preston AR (2012) Hippocampal and ventralmedial prefrontal activation during retrieval-mediated learning supportsnovel inference. Neuron 75:168 –179.


Temporal Dynamics of Memory-guided Cognitive Control and Generalization of Control via Overlapping Associative MemoriesIntroductionMaterials and MethodsResultsDiscussionReferences

temporaldynamicsofmemory-guidedcognitivecontrol ... · these findings were cross-validated using...

Documents